Multi-Channel Multi-Scale Convolution Attention Variational Autoencoder (MCA-VAE): An Interpretable Anomaly Detection...

Area of Science:

• Industrial IoT and Machine Learning
• Time Series Analysis
• Deep Learning for Anomaly Detection

Background:

• Increasing industrial risks necessitate accurate and interpretable anomaly detection.
• Current neural networks struggle to capture both temporal and relational features in industrial data.
• Existing methods that consider both feature types often lack interpretability.

Purpose of the Study:

• To propose a high-precision, interpretable anomaly detection algorithm for industrial applications.
• To enhance the ability of anomaly detection systems to pinpoint specific faulty equipment.
• To improve the restoration of regular operations in abnormal equipment.

Main Methods:

• Utilized a multi-scale local weight-sharing convolutional neural network for temporal feature extraction.
• Employed multiple attention heads to capture inter-dimensional relationships.
• Developed a variational autoencoder (VAE) with an optimization method for anomaly detection via reconstruction errors.
• Leveraged VAE probability distributions for interpretable anomaly scoring.

Main Results:

• Achieved superior performance in anomaly detection, evidenced by high F1 score (0.905) and AUC value (0.982).
• Outperformed the baseline Transformer with a Discriminator for Anomaly Detection (TDAD) by 4% in F1 score.
• Demonstrated accurate anomaly interpretation capabilities, aiding technicians in identifying root causes.

Conclusions:

• The proposed VAE-based algorithm offers a significant advancement in industrial anomaly detection accuracy and interpretability.
• The method effectively extracts complex temporal and relational features from multi-dimensional time series data.
• This approach provides valuable insights for diagnosing and resolving equipment anomalies in factories.